> My first thought would on Loongson would be softfloat bug. > But I am not 100% sure it (still) uses softfloat.
Loongson has never used softfloat. But all mips ports embed the softfloat code in order to perform tricky computations the FPU won't perform (mostly edge cases involving denormals and infinities). So in this case, either the computation gets entirely done in hardware, but the hardware yields a wrong results, and there won't likely be a fix for this, or it asks the kernel for assistance, and there might be a sign bug in there.
